Automated dispersion curve picking using multi-attribute convolutional-neural-network based machine learning

Surface wave dispersion curves are useful to characterize shallow subsurface structures while accurately picking them is typically laborious. To make these approaches more efficient and practical, it is important to automate the picking process. We propose a convolutional neural network (CNN) based ML method to automatically pick multimode surface wave dispersion curves. We modify the typical U-net architecture to convert the conventional 2-D image segmentation problem into direct multimode curve fitting and subsequent picking. A variety of attributes of the data amplitude (A) in the (f, k) domain, such as frequency (F), wavenumber (K), maximum coherency (Coh) and Power weighted amplitude (Pwa), are combined to constrain the picking more accurately than a single attribute does. The effects of two different loss functions on the final picking results are compared; the one that combines conventional wavenumber residuals and curve slope residuals produces more continuous curves. Pre-training the network with synthetic data, and thus using transfer learning, improves the efficiency of the algorithm when the data set is large. To determine the frequency band of each dispersive mode (effective frequency band) in the picked curves, the CNN outputs are post-processed by using measurements such as long/short moving average ratios of squared picked wavenumbers, posterior uncertainty of picked wavenumbers and wavenumbers in the picked curves or the power attribute. We demonstrate the effectiveness of this automatic picking by applying it to a 2-D line and a 3-D subset from a field ocean bottom node data set, where the fundamental and first higher modes of Scholte waves are accurately picked.

Automated summary

In this paper, a CNN-based machine learning method is proposed to automatically pick multimode surface wave dispersion curves. By modifying the U-net architecture and combining multiple attributes for picking constraint, as well as using a comprehensive loss function and pre-training algorithm, the efficiency and accuracy of the algorithm are improved.